Experts Find Flaw in Replicate AI Service Exposing Customers' Models and Data

Researchers disclosed a critical flaw in AI-as-a-service provider Replicate in which malicious models packaged via the Cog tool could achieve remote code execution and abuse a centralized Redis queue to perform cross-tenant attacks, potentially exposing private models, prompts, and PII; Wiz demonstrated the technique, Replicate patched the issue after responsible disclosure in January 2024, and there is no evidence of active exploitation.